[AINews] Reflection Beam - 501B-A23B American Open Model

What it claims

Reflection AI's Beam is a text-only mixture-of-experts model with 501B total and 23B active parameters, trained from scratch in the US for coding, agentic and scientific work. Full weights are due this month under Apache 2.0. Reflection's headline claims are:

• 80.9 on SWE-bench Verified.
• 3–4x the inference efficiency of GLM 5.2.

The newsletter's own framing is less flattering than the summary suggests. Reflection benchmarks Beam against Inkling, Nemotron and GLM 5.2. The current leaders (GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash) are "generally ahead." Beam's main draw is that it is a US-trained open model, not a state-of-the-art one.

How it works

The page gives only partial details, drawn from team posts and outside analysis:

• Pretraining data: 23.8T tokens. Part of this comes from an OCR pipeline run over hundreds of millions of PDFs.
• Post-training: a stable RL and on-policy distillation (OPD) run on about 10K GB300s, with more than 100M rollouts across about 1M tasks.
• Schedule: four weeks of pretraining and four weeks of RL, each on about 10,500 GB300s.
• Architecture (outside reading, not confirmed by Reflection): Elie Bakouch reads it as 3:1 interleaved global and sliding-window attention.
• Compute (outside reading): Teortaxes calls Beam an iso-FLOP replication of DeepSeek V3. He infers about 1.3B RL sandboxes over the 4 weeks, with up to 170K running at once.

Evidence

Beam:

• SWE-bench Verified: 80.9, self-reported. No harness or setting is given.
• Pretraining efficiency: Bakouch estimates only about 12% BF16 MFU. He also reports better held-out code perplexity than DSv4.
• Positioning: observers place Beam around GLM-5.2 level, and below DSv4 Flash on some benchmarks. Nathan Lambert groups it with Nvidia and Thinking Machines as strong US releases that still trail their Chinese counterparts.
• Token efficiency: Artificial Analysis, which has early access, expects Beam to be among the most token-efficient open models for its intelligence level.
• Cost (Axios): Reflection pays $150M/month for Colossus compute and has a separate $1B Nebius deal.

Other items with numbers the summary omits:

• HF multi-harness RL:
• The same weights score 62% under Mini-SWE-Agent and 33% under Claude Code. The harness alone moves results by about 2x.
• Training LFM2.5-2.6B across 4 harnesses lifts first-attempt solves from 42% to 54%.
• A tool-call bonus cuts tool calls by 31%.
• SFT on 3,189 rollouts plateaus at 47.5%.
• SemiAnalysis subscription value:
• Claude plans deliver 5x+ more API-equivalent value than OpenAI plans, measured by per-model, per-token-type credit cost rather than list API prices.
• Adjusting for task cost narrows Claude's edge to 1.3–2.9x.
• Aleph Alpha Kolibri:
• 78B total / 3.46B active, Apache 2.0, built for German and English.
• Self-reported scores: 96.9% AIME 2025, 84.3% GPQA Diamond, 66.4% SWE-Bench Verified.
• NVIDIA mid-harness verification: a GPT-5.6 Sol verifier choosing among 8 candidate shell commands lifts TerminalBench-Lite Pass@1 from 50% to 68%. Weak verifiers add little.
• Google VeriHarness: +6.2 points with Gemini 3.5 Flash and +6.4 with Opus 4.8. About 26K rollouts are released.
• UT Austin context compression (~35K runs):
• Compression that uses a third of the tokens can be 20–80% slower than full context.
• Threshold triggers beat step triggers.
• The best policy varies by model.
• CorpusMap: precomputed entity pages for document collections raise answer quality by 6.4–11.7 points and cut input tokens by 34–57%.
• OpenAI speed change: default speed for GPT-6 Astra and GPT-6.1 Sol rose from about 30 to about 50 TPS. This covers all subscription surfaces and Sign in with ChatGPT partners: OpenCode, Pi, Amp and Devin.

How to use it

• Beam: weights are not out yet. They are promised this month under Apache 2.0, along with a tech report and OSS integrations. No download path, serving recipe or hardware guidance is given.
• Upstage Solar Mini 4: 35B / 3B active with 512K context. It is free on Nous Portal for two weeks.
• Hugging Face harness proxy:
• It speaks the OpenAI Chat, OpenAI Responses, Anthropic and Gemini formats.
• It forwards calls to vLLM and records exact token IDs and logprobs for TRL, so unmodified harnesses become RL environments.
• RL environments are now hosted and versioned on the HF Hub, like datasets.
• Cognition's Agent Memory Repo: the memory format is open-sourced and backed by git and markdown.

Limits

• Beam: every number is either self-reported or a third-party inference made before release.
• There is no tech report yet.
• The 12% MFU, the attention layout and the sandbox counts are outside estimates.
• Reflection's chosen baselines exclude the current open leaders.
• Kolibri: the dataset is unreleased, and its agentic evals sit well below Qwen.
• HF harness RL: the results come from one task family and one seed.
• Unverified claims: the figure that Anthropic spends 42% of inference compute on subscriptions is flagged as unverified. The Microsoft and Meta Claude spend figures are "reported" by The Information.
• Format: the page is a link roundup. Most claims come from tweets, and the detail stops at what is quoted here.